Face Detection Algorithm With Multi-scale And Multi-task Based On Region Convolution

With the boom of deep learning,the Convolutional Neural Network(CNN),as an important branch of artificial neural networks in deep learning,has made great breakthroughs in many fields of computer vision.Especially in the direction of object detection,the performance has been constantly improved because of its excellent characteristics of extracting image features automatically.As a special case of object detection,face detection is also playing a key role in many face applications.However,the detection results will be affected by some factors such as background,lighting,occlusion and expression gestures in the wild,which makes face detection be the challenge to be solved.Therefore,the face detection algorithm based on CNN,especially the general object detection method based algorithm,has been continuously improved,and has achieved good results to effectively addressing the problem of low accuracy in the wild.Based on the general object detection method(i.e.Region-based convolutional neural network,R-CNN),in this thesis,a multi-scale deformable convolutional neural network is designed for multi-scale face detection.R-CNN is a general object detection structure and it belongs to a two-step detector which consists of a feature extraction module,a region proposal module,and a classification regression module.Considering the balance between speed and accuracy,Region-based Fully Convolutional Network(R-FCN)is used as the basic framework of face detection,and some recent achievements is combined including(1)The last two layers of pre-trained CNN are replaced by deformable convolution structure,which can increases the receptive field,that can effectively deal with geometric deformations such as face scale,posture and occlusion.(2)Feature pyramids network is used to solve multi-scale face detection problems.(3)The cross entropy loss is replaced by Focal Loss,which can deal with the imbalance of positive sample and negative sample.The experimental results show that the proposed face detection framework has good performance in multiple face detection benchmarks and is superior to many existing methods.Based on the detection,face attribute recognition further processes and recognizes the face information,which belongs to the multi-label problem.Its purpose is to identify the attributes of the detected faces,such as key points,gender,and posture,etc.Based on the face detection model,the CNN pre-training model extracts shared facial features,and the classification and regression module is divided into several branches,each of which contains features specific for each task,so as to integrate multitasking into a single model.Each task has it`s own feature extract branches and regressions or discriminators.The experimental results show that the proposed multi-scale and multi-task face detection model can achieve good results in both face detection and face attribute recognition,which is better than some existing methods.